Tsimerman Left Toronto for OpenAI. The Real Math-to-AI Pool Is 11.

On July 23, 2026, hours after the International Mathematical Union named him a Fields Medalist for his work on the André-Oort conjecture, Jacob Tsimerman announced he was taking leave from the University of Toronto to join OpenAI's safety team. He is the second Fields Medalist in a year to move toward a frontier lab. Every recruiter at every frontier lab is now hunting the same pool, and that pool is much smaller than the coverage implies.

How thin is the sourceable math-to-AI pool, really?

Eleven people. That is the entire global set of profiles combining mathematician-track credentials with an "AI safety" signal in the headline, according to Refolk's index. Not eleven thousand. Not eleven hundred. Eleven.

That number is the actual order of magnitude every AI lab recruiter is now competing on. Refolk's index also shows 227 profiles worldwide who list Number Theory as a skill and mention AI in their headline, and 930 profiles with the title "Mathematician" or "Research Mathematician" across the US, UK, and Canada combined. The Bureau of Labor Statistics counts roughly 3,500 employed mathematicians in the US at a median salary of $112,110. Stack the numbers next to each other and the story stops being about a hiring boom and starts being about a bidding war on a small, mostly known list.

11

The bottom row is the one to sit with. Three tenths of one percent of employed US mathematicians have publicly aligned themselves with AI safety work in a way a sourcing tool can see. Tsimerman just made that pool one more prestigious and one less available.

Why the "Mathematician" job title is a trap

Filtering by the word "Mathematician" gets you the wrong people. Of the 930 US, UK, and Canada profiles with "Mathematician" or "Research Mathematician" in their title in Refolk's index, the largest single employers are the US Army (8), EY (4), and US Air Force (4). These are defense cryptographers, actuarial consultants, and operations-research staffers, not the pure-math researchers OpenAI, Anthropic, DeepMind, and Math, Inc. are trying to poach.

The mismatch is structural. Academic mathematicians almost never call themselves "Mathematician" on LinkedIn. They call themselves:

• Assistant Professor, Associate Professor, Professor
• Postdoctoral Fellow, Research Fellow, CLE Moore Instructor
• PhD Candidate, Visiting Scholar
• Member (at IAS Princeton or MSRI Berkeley)

None of those title strings retrieve the target population if you search for "Mathematician." A recruiter running a boolean pass on job title will pull an Army cryptographer and miss every Fields Medal shortlist candidate in the country. This is the exact gap Refolk closes: you describe the person in plain English ("research mathematicians in number theory or arithmetic geometry who have started publishing or posting about AI") and get a ranked shortlist built from GitHub, LinkedIn, and open-web signal instead of a keyword match on job title.

What the real sourcing signals look like

Titles fail. What works is a small set of high-signal artifacts that map cleanly onto a real person. Here are the ones that produced verifiable names in the last six months:

1. The Leiden Manifesto on AI and Mathematics (June 2, 2026). Sixteen researchers from 15 universities, official IMU support, signed by Terence Tao and Peter Scholze. This is a curated list of AI-engaged mathematicians handed to you in public.
2. The Stanford Future of Mathematics symposium (early May 2026). Three Fields Medalists, OpenAI and DeepMind researchers, and dozens of mathematicians and computer scientists in one room.
3. The IPAM at UCLA convenings, including the Terence Tao and Mark Chen fireside. Direct evidence OpenAI is systematically farming this community and a documented list of who showed up.
4. arXiv activity in math.NT, math.AG, and cs.LG in the same author profile. The double-listing is the signal.
5. Lean and Mathlib commits. Formal verification is where mathematicians and AI researchers touch code together. Math, Inc.'s Gauss autoformalization model formally verified Maryna Viazovska's 2022 Fields-winning proofs in February 2026. Anyone in that verification pipeline is on the list.

None of these signals live in a standard ATS. None of them appear when you search LinkedIn for "PhD Mathematics." Signed manifestos and workshop invite lists are the highest-signal sourcing artifacts in this category right now, and the first-mover advantage on them is measured in weeks, not quarters.

Filtering by the word Mathematician gets you Army cryptographers. Filtering by Mathlib commits gets you the next Tsimerman.

The competitor set is bigger than OpenAI, Anthropic, and DeepMind

If you are recruiting mathematicians for AI and you think you are competing with the Big Three, you have already lost candidates. The real competitor set for the math-to-AI pool includes at least five other named buyers, each active in the last six months:

• Math, Inc. built the Gauss autoformalization model and formally verified Viazovska's Fields Medal proofs. They are hiring the exact talent every safety team wants.
• Noether Research and Stormglass AI Safety Institute show up as top employers inside the 11-profile intersection in Refolk's index.
• PIBBSS (Principles of Intelligent Behaviour in Biological and Social Systems) recruits from the same pool with a fellowship structure.
• AI Safety Institute (UK) also appears in the intersection set.
• Axiom is an emerging specialist lab in the same category.

The mistake most founders make is assuming OpenAI-scale comp is the winning offer. It isn't, because the candidates aren't optimizing for comp. Tsimerman is taking leave from Toronto and keeping his faculty post. That is the emerging deal structure across this segment: fellowship-style secondments of 12 to 24 months with the tenure seat preserved. Comp still matters, but the offer that wins is the one that lets a prestige-anchored mathematician tell colleagues he "took a leave" rather than "left academia." Full-time offers from unknown labs lose to leave packages from named ones.

Why this happened now, and why it accelerates

The 2026 shift is downstream of three concrete results in a nine-month window. OpenAI disproved the unit distance conjecture in May 2026, a result Quanta described as good enough for top mathematics journals. Math, Inc.'s Gauss formally verified Viazovska's Fields-winning proofs in February 2026. And Terence Tao, speaking at the same Philadelphia congress where Tsimerman won, described mathematics as entering a turbulent period he framed as a crisis in the discipline's foundations and its working values.

Tsimerman himself, at 38, told the press conference AI would soon outperform human mathematicians and could accelerate the field 100 times. When the most decorated 38-year-old in your discipline says that in public, the fence-sitters move. The LessWrong write-up put it bluntly: top mathematicians on par with Critch, Christiano, and Steinhardt are going to frontier labs without even having considered x-risk arguments first.

The mechanism is prestige laundering. A Fields Medal moving to OpenAI legitimizes the move for the next tier, which legitimizes it for the tier after that. Recruiters who wait for the pool to grow are misreading it. The pool of people who publicly signal AI interest will grow, but the pool who will actually respond to a cold outreach in the next 90 days is the same 11 profiles, plus the ~40 living Fields Medalists, plus whoever your competitors already have warm intros to.

~0.3%

The sourcing playbook that actually works on this pool

Stop starting from titles. Start from artifacts. Here is the sequence that produces the shortest useful list for this segment:

1. Pull the manifesto and workshop rosters first. Leiden Manifesto signatories, Stanford Future of Mathematics attendees, IPAM convening participants. This is 100 to 200 named people, most reachable within two email hops.
2. Cross-reference with arXiv. Any author with recent papers in both a pure math category (math.NT, math.AG, math.CT) and a machine learning category (cs.LG, cs.AI, stat.ML) inside 24 months is a high-signal candidate.
3. Add Mathlib and Lean contributors. Formal verification is the technical bridge; commit history is public.
4. Layer in the "taking leave" filter. Faculty with recent visiting positions at frontier labs, IAS members with industry co-authors, postdocs with lab affiliations in their bio.
5. Only then look at LinkedIn. Use it for reachout paths, not for discovery. The 930-profile "Mathematician" search is where you verify, not where you hunt.

This is the workflow Refolk was built for. Instead of running five separate searches across arXiv, GitHub, Mathlib, LinkedIn, and open web and deduping by hand, you describe the target person in one paragraph and Refolk returns the ranked overlap. For a category this thin, the difference between an eleven-person shortlist and a nine-hundred-person miss list is the difference between filling the role in a quarter and losing the candidate to Math, Inc. in a week.

What to do this week if you are recruiting for this pool

Three concrete moves, in order:

• Save the Leiden Manifesto signatory list and the Stanford May 2026 attendee list as source lists in your ATS. These are 100 to 200 names combined. They are the warmest cohort in the category.
• Draft a leave-structured offer, not a full-time offer. 18-month research fellowship, faculty seat preservation clause, publication rights, named collaborator. This is the format that beats a straight comp package on this pool.
• Run the plain-English query on Refolk for "research mathematicians pivoting to AI safety." The output is small enough to work through candidate by candidate. That is the point.

The next Tsimerman is on one of those lists already. You are not going to find him by searching for the word "Mathematician."

FAQ

How many mathematicians are actually available to hire into AI labs right now?

Refolk's index shows 11 profiles worldwide at the exact intersection of mathematician-track credentials and an "AI safety" headline signal, and 227 profiles globally who list Number Theory as a skill and mention AI in their headline. The broader "Mathematician"-titled pool across the US, UK, and Canada is 930, but that population is dominated by US Army, EY, and US Air Force employers, not the academic pure-math talent frontier labs want. The practically reachable pool is in the low hundreds at most.

Who competes with OpenAI for this talent?

Anthropic and DeepMind are the obvious ones, but the real competitor set includes Math, Inc. (which built the Gauss autoformalization model and formally verified Viazovska's Fields-winning proofs), Noether Research, PIBBSS, Stormglass AI Safety Institute, Axiom, and the UK AI Safety Institute. Several of these appear as top employers inside Refolk's 11-profile intersection set, which means they are already winning candidates from the exact pool larger labs are chasing.

Why doesn't a LinkedIn search for "Mathematician" work for this?

Because academic mathematicians almost never use "Mathematician" as their job title. They use Professor, Postdoctoral Fellow, PhD Candidate, or Member. The 930 people who do use the "Mathematician" title in the US, UK, and Canada are concentrated at the US Army, EY, and US Air Force, meaning defense and consulting rather than research. Sourcing this pool requires arXiv, Mathlib and Lean commits, and workshop attendance signals, not job title filters.

What kind of offer actually wins in this category?

A leave-structured research fellowship, not a full-time hire. Tsimerman is taking leave from the University of Toronto while keeping his faculty post, and this is emerging as the standard deal shape. Frontier labs that offer 12-to-24-month secondments with faculty seat preservation and publication rights consistently beat higher-comp full-time offers, because prestige-anchored mathematicians are not willing to fully cut the tenure cord for any amount of money.